Beyond ChatGPT: Building Real AI Workflows for Finance & Accounting

In this MizuFlow.ai Foundation of Finance episode, Sung Lee, CFA, CPA, CA, explains how organizations transform Large Language Models (LLMs) into practical enterprise intelligence systems.

While modern LLMs are incredibly powerful, they have important limitations:

• no access to private company data
• knowledge cutoffs
• hallucination risk
• inability to perform real-world actions

This session explores how technologies such as:

• Retrieval Augmented Generation (RAG)
• Tool Calling
• API Integration
• Structured Outputs
• Agentic Workflows

enable AI systems to move beyond conversation and become active participants in enterprise operations.

🧠 What This Video Covers
Why LLMs Alone Are Not Enough

A standalone LLM only knows what it learned during training.

It does not automatically know:

• your ERP data
• your contracts
• your financial statements
• your policies
• your customer records

This creates significant limitations for enterprise applications.

Retrieval Augmented Generation (RAG)

The module introduces one of the most important enterprise AI architectures:

Retrieval Augmented Generation (RAG)

Instead of relying solely on training data, the AI retrieves relevant information from:

• accounting policies
• financial statements
• contracts
• audit documentation
• knowledge bases
• enterprise document repositories

before generating a response.

This dramatically improves:

• accuracy
• relevance
• traceability
• enterprise usefulness

Tool Calling

The next evolution involves:

Tool Calling

Rather than simply answering questions, AI systems can invoke external tools.

Examples include:

• querying databases
• retrieving reports
• calculating forecasts
• running Python code
• searching document repositories

This allows AI to interact with business systems in real time.

API Integrations

Through APIs, AI can connect directly with:

• SAP
• Oracle
• Salesforce
• Power BI
• accounting systems
• banking platforms
• custom applications

This transforms AI from:

Information Provider

into:

Business Process Participant
Structured Outputs (JSON)

A major challenge in enterprise automation is reliability.

The session explains how:

Structured Outputs

allow AI systems to produce predictable results.

Instead of generating free-form text, AI returns:

• JSON objects
• predefined schemas
• structured fields

This enables software systems to safely consume AI outputs.

Example:

{
"vendor": "ABC Supplier",
"invoice_amount": 12500,
"approval_required": true
}

This becomes critical for enterprise workflows.

Agentic Workflows

The module then brings everything together.

Modern AI agents combine:

• reasoning
• retrieval
• tools
• APIs
• structured outputs

into integrated business workflows.

Example:

Question
↓
RAG Search
↓
Reasoning
↓
Tool Calling
↓
ERP Update
↓
Human Review

This creates intelligent automation systems capable of completing complex tasks.

Finance & Accounting Applications

Examples include:

Financial Reporting Agents

• retrieve supporting schedules
• draft financial statement notes
• generate variance commentary

AP Automation Agents

• read invoices
• extract fields
• validate vendors
• route approvals
• update ERP systems

FP&A Agents

• gather assumptions
• run forecasts
• compare scenarios
• generate management commentary

Tax & Research Agents

• retrieve authoritative guidance
• summarize conclusions
• prepare workpapers

Human Oversight Still Matters

A major theme throughout the session is:

Human-in-the-Loop Governance

Even the most sophisticated agentic systems require:

• approval workflows
• audit trails
• governance controls
• professional judgment

AI should prepare.

Humans should approve.

🚀 Why This Matters

Enterprise AI is no longer just about asking questions.

The future involves systems that can:

• reason
• retrieve information
• call tools
• interact with software
• complete workflows

This is the foundation of:

• AI copilots
• accounting agents
• enterprise automation platforms

🎯 Key Takeaway

LLMs alone are powerful.

But enterprise intelligence emerges when we combine:

✅ RAG
✅ Tool Calling
✅ APIs
✅ Structured Outputs
✅ Human Oversight

This creates agentic systems capable of supporting real business operations.

The future of finance is:

Human Judgment
+
AI Agents
+
Enterprise Data

DISCLAIMER & LIABILITY NOTICE: The content in this video is for educational and informational purposes only. It does not constitute financial, accounting, tax, or legal advice.

No Professional Relationship: Watching this video or interacting in the comments does not create a CPA-Client or fiduciary relationship between you and Sung Lee.

Software & Tools: Any code, software, or tools mentioned (including https://www.google.com/search?q=Katchiflow.com) are provided "as-is" for demonstration and drafting purposes only. Outputs should not be relied upon for tax or statutory reporting without independent verification by a qualified professional.

Видео Beyond ChatGPT: Building Real AI Workflows for Finance & Accounting канала MizuFlow